Evaluation of a fully automated 2-dimensional imaging system for real-time cattle lameness detection using machine learning

Siachos, N ORCID: 0000-0001-7670-4950, Griffiths, BE ORCID: 0000-0002-2698-9561, Wilson, JP ORCID: 0000-0001-7096-8007, Bedford, C ORCID: 0009-0004-3103-0920, Anagnostopoulos, A ORCID: 0000-0002-5193-858X, Neary, JM ORCID: 0000-0001-8438-2234, Smith, RF ORCID: 0000-0003-0944-310X and Oikonomou, G ORCID: 0000-0002-4451-4199 (2025) Evaluation of a fully automated 2-dimensional imaging system for real-time cattle lameness detection using machine learning Journal of Dairy Science, 108 (4). pp. 4206-4224. ISSN 0022-0302, 1525-3198

Text Siachos et al 2025 CattleEye PIIS0022030225000840.pdf - Unspecified Available under License Creative Commons Attribution. Download (2MB) | Preview

Abstract

Early detection and prompt treatment of lame cows are crucial for proactive lameness management. This study aimed to evaluate a fully automated 2-dimensional imaging system for real-time lameness detection using artificial intelligence. Data were collected from 11 dairy farms in the UK Four trained veterinarians performed 42 mobility scoring sessions using a 0–3 4-grade scoring system, with scores 2 and 3 representing lameness. On each session, individual weekly average scores were calculated. This resulted in 40,116 paired human mobility scores (HMS) and weekly average mobility scores generated using artificial intelligence (AIMS) matched to a cow ID. Categorical agreement for the 4-grade scale was estimated by calculating the weighted Cohen's kappa (κ<inf>w</inf>) and Gwet's agreement coefficient (AC<inf>2</inf>), and for the 2-grade scale (nonlame vs. lame) by calculating the percentage agreement (PA), unweighted Cohen's kappa (κ) and Gwet's coefficient (AC<inf>1</inf>). A trained veterinarian recorded the presence and severity of any lesion of 2,515 cows, which also had an AIMS assigned. A subset of 758 cows were also assigned an HMS 1–3 d before trimming. Sensitivity (Se), specificity (Sp), and accuracy (Acc) were calculated to describe the system's and human's ability to detect cows with foot lesions. Additionally, automated mobility scores were retrieved for cows with foot lesion records up to 30 d before trimming. Linear mixed effects models (LMM) were built to assess the association of the lesion status at trimming with the daily scores. The average (mAVG), maximum (mMAX), minimum (mMIN) and the percentage of scores that a cow was identified as lame (mPLS) during the 30 d before foot trimming were calculated and their Se, Sp and Acc in detecting foot lesions were determined. Lastly, longitudinal data were obtained from 143 cows tracking daily scores from 5 to 64 DIM. The association of lesion status at the early lactation routine trim (ELRT) with the daily scores was assessed by fitting LMM. Regarding the 4-grade scale agreement between HMS and AIMS, κ<inf>w</inf> (0.24–0.34) represented fair agreement, whereas AC<inf>2</inf> (0.81–0.93) almost perfect agreement. For the 2-grade scale agreement, PA was consistently above 80%, κ (0.23–0.38) represented fair agreement, and AC<inf>1</inf> (0.76–0.83) showed substantial to almost perfect agreement. The AIMS detected cows bearing severe lesions with Se = 0.53 and Sp = 0.74, whereas the HMS achieved Se = 0.60 and Sp = 0.78. Using optimal thresholds for mAVG, mMAX, mMIN, and mPLS, the system achieved higher Se than HMS. Moreover, cows with severe lesions had increased scores from 23 d before trimming compared with cows with mild and moderate lesions. Longitudinal data showed that cows with severe lesions at ELRT had higher mobility scores during the first 60 DIM compared with those with mild or moderate lesions. Overall, the system's performance was comparable to that of experienced human assessors in detecting lame cows and cows with foot lesions. Finally, its capability to detect mobility changes before the development of severe lesions highlights its potential for early intervention, which could enhance lameness management in dairy herds.

Item Type:	Article
Uncontrolled Keywords:	artificial intelligence, convolutional neural network, foot pathologies, locomotion, mobility
Divisions:	Faculty of Health & Life Sciences Faculty of Health & Life Sciences > Inst. Infection, Vet & Ecological Sciences
Depositing User:	Symplectic Admin
Date Deposited:	06 Mar 2025 15:43
Last Modified:	28 Feb 2026 11:01
DOI:	10.3168/jds.2024-25940
Related Websites:	Author Publisher
URI:	https://livrepository.liverpool.ac.uk/id/eprint/3190703
Disclaimer:	The University of Liverpool is not responsible for content contained on other websites from links within repository metadata. Please contact us if you notice anything that appears incorrect or inappropriate.